Proof of Inference

3. Consensus Mechanism: Proof-of-Inference (PoI)

Traditional computational marketplaces suffer from the "hardware spoofing problem"—malicious actors claiming premium GPU capabilities while delivering inferior performance. NeuroGrid's Proof-of-Inference consensus eliminates this attack vector through cryptographic work verification.

3.1 Challenge-Response Protocol

The Challenge Vector: At pseudo-random intervals (governed by a verifiable random function), the protocol broadcasts a deterministic matrix multiplication benchmark—termed a "HeARTbeat"—to nodes reporting idle capacity. This computational challenge is specifically designed to be:

• Hardware-Dependent: Execution time correlates directly with GPU architecture (memory bandwidth, CUDA core count)

• Unforgeable: Result hashes are deterministic and cannot be pre-computed without actual execution

• Lightweight: Challenge completion consumes <0.1% of node resources, ensuring minimal interference with active rentals

The Response Window: Nodes must return a cryptographically signed proof within a latency threshold calibrated to their declared hardware specifications. For example, an H100 node claiming 80GB HBM3 memory must complete challenges 3.2× faster than a 4090 node, as validated by the protocol's hardware performance database.

Mathematical Verification:

Expected_Latency = f(GPU_Model, Matrix_Size, Memory_Bandwidth)
Actual_Latency = Timestamp_Response - Timestamp_Challenge

Verification_Result = |Actual_Latency - Expected_Latency| < Tolerance_Threshold

The Adjudication Mechanism: Nodes failing verification (either through incorrect results or timeout violations) incur:

• Immediate Suspension: Temporary removal from the active provider pool (30-minute cooldown)

• Reputation Degradation: Exponential decay of the node's Trust Score, reducing visibility in marketplace rankings

• Collateral Slashing: Repeated failures (>3 in 24 hours) trigger partial confiscation of staked $NRG collateral

This cryptoeconomic penalty structure ensures that rational actors maintain honest hardware declarations, as the cost of deception exceeds potential fraudulent gains.